Master Thesis Defense by Svenja Frey

Title: TRACE -- Toward Reconstructing causal dynamics of Atmospheric CO2 Evolution under Dansgaard–Oeschger Events using Machine Learning Approaches

Abstract:

Many different research approaches have been taken towards understanding the underlying causes of atmospheric CO2 variations during Dansgaard-Oeschger events. Simultaneously, machine learning methods are on the rise and open up new possibilities. This thesis aims to investigate selected machine learning and causal discovery approaches by analyzing their success in reproducing known causal structures in climate model data. We use long integrations of a last glacial climate setup of the Community Earth System Model with freely oscillating Dansgaard-Oeschger dynamics and full biogeochemistry. Extensive research has been carried out in analyzing the carbon cycle during DO-Events of this configuration, which the methods will be evaluated against. We show that the popular causal discovery routine for climate data, PCMCI+, struggles to reproduce the known underlying causal structure in the data. Furthermore, its sensitivity to algorithm parameters introduces uncertainty due to large differences in the found causal graph. We then switch to Convolutional Neural Networks (CNNs) and a custom causal discovery algorithm is presented combining CNNs with flexible lags between variables and Pearl's causality framework, bridging the gap between causality and machine learning. This approach is successful in identifying the latent causal graph known from the data: AMOC is the driver of atmospheric CO2 changes during DO-Events, not Southern Hemisphere processes. To gain an alternative angle, we utilize layer-wise relevance propagation (LRP), an explainable AI (XAI) approach, which shines light on the decision processes that lead a CNN in its prediction. While this technique shows some success in identifying the right timing and attribution of feature contributions, it ultimately fails to make the jump from correlation to causation. At its core, this thesis highlights the importance of embedding machine learning methodologies firmly within a causal framework — precisely because it prompts the question: how causal is machine learning, really?

Supervisor: Marcus Jochum
Censor: Ulrik Lumborg